Research And Design Of Detection Algorithm For Floating Objects On Water Surface

With industrial progress and social development,the problem of water surface floating matter pollution seriously affects human production and life,and real-time detection of water quality is an important part of water quality management and floating matter pollution prevention.Due to the complexity of the water surface environment,the water surface image has the characteristics of uneven illumination,easy to be polluted by noise,susceptible to weather,etc.,making the detection of water surface floaters has certain special characteristics,the traditional target detection algorithm in the external environment or noise interference and other circumstances feature extraction capacity is limited.In this paper,we take water surface floating objects(water hyacinth,cyanobacteria,floating pigs,ships,etc.as detection objects and construct a target detection model for water surface floating objects to solve the problem of real-time detection of water surface floating objects in real scenarios,which is of practical significance for promoting the development of water surface environment detection in the direction of intelligence.（1）To address the problems of poor sample quality and large differences in existing data sets,the actual situation of the application scenario is analyzed,and the output of the real-time image from the camera is collected as samples of the water surface floating object data set.To solve the problems of a limited number of surface target samples and class imbalance,the single-sample data enhancement is combined with the Borderline SMOTE algorithm to complete the pre-processing of the water surface floater dataset,and finally,the construction of the dataset is completed by image annotation and format conversion,and the dataset is comprehensively evaluated.（2）To address the problems of low performance and poor accuracy of traditional target detection algorithms in the process of water surface environment detection,we compare the detection performance of Faster-RCNN,SSD,and YOLOv5 detection algorithms in the water surface floating object dataset.The principles,network structures,and loss functions of the three classical target detection algorithms are briefly described,the hyperparameter settings of the detection network are completed,the performance of the network model is evaluated using evaluation indexes such as accuracy,recall,regression loss,and classification loss,and finally,YOLOv5,which has the best overall performance,is selected as the base model for subsequent improvements.（3）To improve the detection accuracy of the YOLOv5 target detection algorithm,the CAIA＿YOLOv5 target detection model for floating objects on the water is designed.The sample labels are softened by introducing smoothing labels in the input.introducing coordinate attention mechanism in the backbone network to complete accurate localization and recognition of the target of interest,and designing a sparsely connected residual structure to increase the network perceptual field while reducing the number of network parameters;introducing the PANet structure of adaptive spatial feature fusion algorithm in the neck network to adaptively screen conflicting information;using CIOU＿Loss as the regression loss function to combine the distance information of bounding box and scale information to accelerate the convergence speed of network regression and classification.（4）Using Py Charm as the development environment,combining Redis database,Pytorch deep learning image recognition framework,a PC server,and an HD camera to build a water floating object target detection system,transplanting the trained CA-IA＿YOLOv5 network to the water floating object detection model,conducting overall debugging analysis and practical deployment,and realizing online floating object detection and real-time push.